Player's guide · World models · October 27, 2026

Genie 4, explained: how Google DeepMind's world model draws a game with no game engine

Google DeepMind's Genie 4 looks like a video game you can walk around in. There is no game engine underneath, no level file and no 3D model of the world. Every frame you see is a prediction, drawn by an AI model from the frames before it and the key you just pressed. Announced on October 27, 2026, it adds sound, characters who talk back, and the ability to change the world by typing.

New explainers when AI news breaks: follow @EmergentMind on X or subscribe on YouTube.

The short version

  1. A world model is an AI that predicts what a scene will look like next, given what came before and what you do. Genie 4 does that fast enough to play in real time.
  2. Nothing about the world is stored as a map. If a door is still there when you turn around, it's because the model remembered it.
  3. New in Genie 4: it generates its own audio (music, ambient sound and dialogue), and you can type to change the weather, move the camera or add people while you're inside.
  4. It's a research preview for trusted testers. Google says Google AI Ultra subscribers (in Project Genie) and early access partners of the Gemini API and AI Studio come next.
The same story in under two minutes.

Level 1 What happened

On October 27, 2026, Google DeepMind announced Genie 4, which it calls its "most realistic and interactive world model yet." The announcement post, by Raia Hadsell, Diego Rivas and Hang Qi, describes four things that are new:

  • Photorealism. Genie 4 builds on Gemini. Google says Gemini Omni "anchors simulations in real-world appearance and behavior," and other Gemini models provide the reasoning to keep scenes coherent.
  • Physics and people. DeepMind highlights aerodynamics, collisions, and human characters who move naturally and can be steered in real time.
  • Native audio. "Genie 4 is our first real-time world model to generate audio natively": music, ambient noise and dialogue, with characters who "converse and react to your inputs or events in the world, all in real-time."
  • Live steering. You move with ordinary keyboard controls and type prompts to change the weather, move the camera, or add a person or object.

DeepMind's first post about it went out at 12:10 pm Eastern and passed a million views within about five hours.

Level 2 How it works: one frame at a time

A normal video game stores the world: a map, 3D models, rules for physics, and a renderer that turns all of that into pixels. Genie works the other way around. It has no stored world. It generates the picture directly.

DeepMind described the method when it announced Genie 3 in August 2025: the model is auto-regressive. Each new frame is generated from the description of the world, the frames that came before, and your latest action. Then the new frame becomes part of the history, and it does it again. Genie 3 ran at 720p and 24 frames per second. DeepMind hasn't published resolution or frame rate figures for Genie 4.

Because nothing is stored as a 3D map, consistency is something the model has to learn. DeepMind called it "an emergent capability" for Genie 3, which stayed consistent for "a few minutes" and could remember what it had drawn "as far back as one minute ago."

Try it: a world that only exists while it's remembered

A toy illustration of the idea, not Genie itself. The world below has no map. Each time a spot comes into view, the "model" reuses what it drew there if it can still remember it, and otherwise draws something new. Walk right, then walk back, and see whether the door is still there.

You're standing next to a door. Walk away and come back.

Steps taken: 0 · Spots redrawn from scratch: 0

Genie 4 adds a planner on top. Google says Gemini adapts each typed request to the scene and "steers the world between inputs" to keep it cohesive, and that Gemini 3.8 Flash acts as a guiding agent: it watches the scene, breaks a goal into steps, checks whether each step worked, and replans when the world changes.

Level 3 The limits

DeepMind is direct about them. Genie 4 "can maintain coherent environments over short sessions," but visual details and object permanence "can still shift or drift over time," especially in longer runs. It lists complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency as active research areas. The announcement gives no benchmark numbers.

Google says Genie 4 has safety guardrails like its other models, was evaluated internally before the preview, and that all generated video and audio carry SynthID watermarks.

Level 4 Why it matters beyond games

Google pitches world models as places to test things before doing them for real. DeepMind paired Genie 4 with Gemini 4 Argon in robotics research: Gemini breaks a task into steps, which are fed to Genie as text, and the example shows Apptronik's Apollo humanoid playing checkers in a simulated scene. DeepMind says trusted testers in film, journalism, education, gaming and engineering have used it for previsualization, turning archival photos into simulations, quick game prototypes, and training scenarios.

Level 5 Reactions

Typed "Neogothic brutalist city inhabited by robots and horses," got it in about a minute, then added dragons and a friendly otter character, all in real time. "It isn't flawless."

Ethan Mollick, Wharton professor who studies AI (@emollick), Oct 27

"This was the last project I worked on at Google," and its quality and dynamics are "still so much better than the 3D creations from the latest frontier models."

Ben Poole, former Google research scientist (@poolio), Oct 27

Highlighted steering environments in real time, native audio "from realistic background sounds to characters who can converse in real time," and photorealistic scenes.

Level 6 Timeline

  1. Aug 5, 2025 DeepMind announces Genie 3: real-time worlds at 720p and 24 frames per second, consistent for a few minutes.
  2. Oct 27, 2026 Genie 4 announced as a research preview for trusted testers, with native audio, talking characters and live text steering.
  3. Coming soon Google AI Ultra subscribers in Project Genie, and early access partners of the Gemini API and Google AI Studio (no date given).

Level 7 Questions people are asking

Is this just an AI video?

No. A video model makes a clip you watch. Genie generates the next frame in response to what you do, so you can walk around, turn back, and type changes while it runs.

Is it a game engine, or is there a 3D world underneath?

DeepMind says Genie 3 used no explicit 3D representation, and its consistency emerged from training. Genie 4 builds on that line and on Gemini. What you see is drawn frame by frame.

If I open a door and walk away, will it still be open?

Over short sessions, it should. DeepMind says details and object permanence can still shift or drift over longer runs.

Can I use it now?

Only trusted testers can today. Google says it's coming soon to Google AI Ultra subscribers in Project Genie, and to early access partners of the Gemini API and AI Studio. There's a waitlist on the announcement page.

How can I tell if something was made with it?

Google says all Genie 4 video and audio carry SynthID watermarks, which Google's detection tools can check.

Save point Updates

  • Oct 27, 2026 This page went up. It covers everything known as of October 27, 2026.

Credits Sources

Want the next one? Follow @EmergentMind · YouTube